[forgotpw1123]

The level of indirection is minimal, basically c++ with garbage collection. That's why they're statically typed.

Both perform within a small margin of native code, with the major exceptions of vector instructions that aren't supported by all CPU's

[flukus]

In an artificial benchmark they are pretty similar, but in real world idiomatic code they are very different to c++. Nearly everything is on the heap so you've got pointers everywhere, dynamic data structures are used everywhere, virtual functions (particularly with java) and interfaces. There's is a reason java was so much slower than native code 20 years ago and that gap has been bridged by CPU progress more then compiler progress.

[pjmlp]

You are forgetting the crappy programmers as well.

I lost count how many times I have reviewed code where a for loop is used instead of System.arrayCopy().

Or the times I have seen code that my CS data structures teacher would probably give a plain 0, even if done in C.

[flukus]

It's probably for the best those people aren't writing c/c++ ;)

I'd love to make them code in rust, their inability to write code, to even copy/paste it will save so many maintenance problems.

[forgotpw1123]

This just isn't true. There's many Java ports of C/C++ applications and benchmarks are usually within 30%.

[coldtea]

For the easy tasks. Try a CPU heavy C/C++ app -- like an audio/video editor, number crunching of any kind, etc.

[forgotpw1123]

Like I said, the major exception is SIMD instructions. AV and heavy math are a tiny fraction of applications. For these you can use JNI.

There is precedent for this in Netty where they use a native poll mode driver and everything else Java, and it performs about as fast as pure C code on Techempower benchmarks